Different types of media bias in online news

Franziska Löw

December 04, 2018

Agenda

  1. Introduction

  2. Literature

  3. Data

  4. Methodology

  5. Results

Introduction

Media and Politics

Hypothesis:

  • A biased media reporting in political news may have a profound influence on voter opinions and preferences

  • The concept of media bias encompasses different subtypes (D’Alessio and Allen, 2000): (1) Coverage bias, (2) Tonality bias, (3) Agenda bias

Research Question:

  • Is online media biased?

  • What is the effect on voting behavior?

Methodology:

  • Calculate the different biases using a combination of different text-mining techniques.
    • extensive content analyses of large text data
    • reduces human induced bias and makes research more traceable and comparable
  • Estimate the effect of different bias measures on voting preferences

Literature

Bias and balance

Unbiased reporting:

  • neutral / balanced report that is not strongly slanted in favor of or against any political side. All sides should be equally represented according to some benchmark.

Biased reporting:

  • To what extend does a reporting deviate from that benchmark?

  • What is the benchmark?

3 types of media bias

  • The concept of media bias encompasses different subtypes (D’Alessio and Allen, 2000):
  1. Coverage bias (Quantity: How often do political actors appear in the media?)

  2. Tonality bias (Quality: How are parties evaluated?)

  3. Agenda bias (Quality: Is a party able to present their own political positions in the media?)

Coverage bias

Hypothesis:

  • Voters tend to prefer parties that are more visible in their media repertoire (Eberl et al., 2017)

Benchmark:

  1. The amount of a parties campaign communication (Hopmann et al., 2012)

  2. Standing in polls (Junqué de Fortuny et al., 2012)

  3. Average visibility of all parties in each media outlet (Eberl et al., 2017)

Methodology:

  • Manually coded data

  • Textmining approach (e.g. count number of occurrences)

  • To estimate the effect of the bias on voting preferences…

  • Regional differences in the reach of certain media (DellaVigna and Kaplan, 2006; Enikolopov et al., 2011; Snyder and Strömberg , 2010).
  • Historial data (Dewenter et al., 2018)

Tonality bias

Hypothesis:

  • Positive or negative aspects of an object highlighted in the media affects the attitudes towards a certain subject (de Vreese & Boomgaarden, 2003; Druckman and Parkin, 2005; Eberl et al., 2017; Junqué de Fortuny et al., 2012)

Benchmark:

  • Average sentiment value of other parties (Junqué de Fortuny et al., 2012) in each media outlet (Eberl et al., 2017)

Methodology: - Manually coded data (Eberl et al., 2017)

  • Textmining approach (e.g. dictionary based methods, Junqué de Fortuny et al., 2012)

Coverage & tonality bias (results)

  • Studies investigating both visibility and tonality bias conclude that these biases are not necessarily consistent in their effects, with tonality bias identified as having a greater impact (Boomgaarden & Semetko, 2012; Norris, Curtice, Sanders, Scammell, & Semetko, 1999; Eberl et al. ,2017).

Agenda bias

Hypothesis:

  • Greater visibility of the political content of a party can have a positive impact on attitudes towards that party. (Benewick et al., 1969; Eberl et al., 2017)

Benchmark:

  • Parties’ campaign communication as approximation of the potential universe of news stories (Eberl et al., 2017)

Methodology:

  • Manually coded data (Eberl et al., 2017)

Results:

  • Voters evaluate parties more favorably if those parties addressed their own favored topics more prominently in media coverage (Eberl et al., 2017)

Methodology

Data

Online news articles

  • n = 11.880
  • time: 01.06.2017 - 01.03.2018

Visibility

A party \(p\) is treated as visible in an article of newspaper \(s\) if the party itself is mentioned in an article (if an article contains the word “SPD”, “CDU”/“CSU/Union”,“FDP”,“Grüne”,“AfD” or “Linke”)

\[ \text{visibility}(p,s) = \frac{\text{#Articles}_{p,s}}{\text{#Articles}_s} \]

{width:80%}

Visibility bias

There is a bias in a medium when a political party is overrepresented compared to other parties and other outlets

  • Benchmark: Average visibility of all parties in each media outlet during the period of analysis

  • Visibility bias: Deviation of each party’s specific visibility from the average visibility of all other parties in that outlet.

Tonality

  • Dictionary-based method to measure the tone (or sentiment) of a document.

  • Dictionary is a list of words associated with positive and negative polarity weighted within the interval of \([-1; 1]\)

word value
1 räuber -0.3508
2 zusammenhalt 0.1947
3 desaster -0.3413
4 vernichtung -0.4883
5 weglassen -0.3362
6 unzumutbar -0.34
7 erholen 0.2357
8 niederschlagen -0.1871
9 abstürzen -0.4697
10 heftig -0.1819
  • The sentiment score for each party in an article is calculated from each word that occurs in a window of two sentences before and two sentences after the occurence of that political party.

Tonality bias

Agenda bias:

Agenda bias refers to the extent to which political actors appear in the public domain in conjunction with the topics they wish to emphasize.

Benchmark: Policy issues addressed in party press releases as an approximation of the potential universe of news stories. .

How to find out latent topics in an article?

Topic Model

Credits: Christine Doig

The intuition behind LDA

Credits: Blei (2012)

LDA as a graphical model

  • \(N =\) collection of words within a document.
  • \(D =\) collection of documents within a corpus.
  • observed: word in a document \(w_{d,n}\)
  • fixed: mixture components (number of topics \(K\) & vocabulary)
  • hidden: mixture proportions (per-document topic proportions \(\theta_d\) & word-topic distribution \(\beta_k\))

Generative process

  • \(K\): choose the number of topics

    • \(K=3\)
  • \(\theta_d\): for each document \(d\), choose a distribution over topics;

    • \(\theta_d\) ~ Dirichlet(\(\alpha\))

  • \(z_{d,n}|\theta_d\): according to \(\theta_d\), assign a topic \(z_{d,n}\) for the \(n^{th}\) word;

    • \(K=Topic 1\)

Generative process (contd.)

  • \(w_{d,n}|z_{d,n},\beta,\theta\): choose a term from that topic according to \(\beta_k\)

    • \(\beta_k\) ~ Dirichlet(\(\eta\))

  • \(N\): repead this process for all \(n\) word-positions in the document.

  • \(D\): conduct this process for all \(d\) documents in the corpus

Strucutral Topic Model (Roberts et. al. (2016))

Allows to include covariates into a topic model:

  1. Topic Prevalence: Attributes that affect the likelihood of discussing topic \(k\)

  2. Topic Content: Attributes that affect the likelihood of including term \(w\) overall, and of including it within topic \(k\)

  • I assume that the topic prevalence depends on the source (Bild.de, FOCUS ONLINE, FDP, …) and the topic content differs between the type of sources (e.g. if the document is a party press releases or a news articles).

Model Results

Label topics

Joint label News articles Press releases
Topic 47 : staatsanwaltschaft polizei erklärt peter staatsanwaltschaft polizei ermittlungen cdu erklärt peter netzdg steudtner
Topic 44 : bundeswehr leyen bundeswehr bundesregierung bundeswehr leyen soldaten nato bundeswehr bundesregierung syrien nato
Topic 27 : spahn cdu gabriel macron spahn cdu jen heimat gabriel macron französischen deutsch
Topic 1 : europa eu eu europäischen europa eu deutschland europäischen eu europäischen europäisch button
Topic 2 : spd schulz groko traurig spd schulz gabriel partei groko traurig verzicht übernimmt
Topic 25 : stiftung deutschland deutschland forschung stiftung deutschland menschen osten deutschland forschung menschen wirtschaft
Topic 42 : abschiebungen deutschland pflege gesetz abschiebungen deutschland abschiebung afghanistan pflege gesetz erklärt jahr
Topic 4 : wahl bundestagswahl gewählt europarat wahl bundestagswahl mitglied parteien gewählt europarat parlamentarischen vorsitzenden
Topic 35 : welt sonntag bund bildung welt sonntag menschen meinung bund bildung länder schulen
Topic 30 : afd höcke meuthen bundesvorstand afd höcke partei gauland meuthen bundesvorstand partei verfassungsschutz

Topic distribution

For each document, we have a distribution over all topics, e.g.:

Expected topic frequency

Measure Agendas

Agendas were measured in terms of percentage distributions across the 50 topics.